Persistence in Windows
SoK: Systematization, Detection, and Hunting of Windows Malware Persistence Techniques
ASIA CCS '26: Proceedings of the ACM Asia Conference on Computer and Communications Security June 2026
- The AV-TEST Institute records over 250K new malware samples every day.
- https://www.av-test.org/en/statistics/malware/
-
We present the largest documented set of persistence techniques (72 in total)
- 48,873 detonated Windows malware samples.
Prevalence
- Only 55.2% of samples are persistent.
- Contrary to the general expectation that the vast majority of active malware requires persistence.
- Absence of running malware does not prove that no execution occurred—malware may exfiltrate data and delete itself without being resident
How
- Over 30% employ two or more persistence techniques
- Removing a single vector may not fully eradicate the malware.
- We identified 42 distinct methods out of 72 documented, highlighting a long-tail distribution. The top 10 techniques cover many families, but full detection requires attention to rarer methods.
- Malware families often switch persistence techniques over time.
- Knowing the family and its typical behavior doesn’t guarantee knowledge of the specific technique used.
- Only 43.2% aim for execution at boot.
- Although persistence is often equated with boot-time execution, many samples prefer to trigger on system events, schedules, or user actions.
Evasion
- Although monitoring Windows API calls is effective for most techniques we documented, attackers can exploit less commonly monitored Windows subsystems to achieve the same effect more stealthily.
- Another common evasion strategy is the obfuscation of Registry and file system paths.
Detection
- While persistence detection is generally considered a solved problem, we reveal that industry-standard persistence detection tools produce a significant number of false positives and false negatives.
- Our analysis of state-of-the-art tools reveals that:
- Capa 72.5% false positives, 42.1% false negative rate (FNR). Triage unknown files, guide reverse engineering, and hunt across a corpus for novel malware. In contrast to YARA, which primarily searches for sequences of bytes, capa rules describe features at the code-level.
- Autoruns 24.9% FNR. This utility, which has the most comprehensive knowledge of auto-starting locations of any startup monitor, shows you what programs are configured to run during system bootup or login…Autoruns reports Explorer shell extensions, toolbars, browser helper objects, Winlogon notifications, auto-start services, and much more.
- Persistence Sniper a 34.9% FNR. Powershell module that can be used by Blue Teams, Incident Responders and System Administrators to hunt persistences implanted in Windows machines.